Microsoft AI-103: Azure AI Apps and Agents Developer Certification Guide

Generative AI has moved from demos to products, and Microsoft's AI-103 exam is its way of recognising developers who can build those products properly on Azure. Passing it earns Microsoft Certified: Azure AI Apps and Agents Developer Associate. The exam covers far more than calling a model: planning and managing AI solutions, building agents, grounding answers in data, analysing text and images, extracting information from documents, and applying safety and evaluation throughout.

This guide explains who AI-103 is for, the exam facts, the skills measured and what sits inside each area, what a successful candidate looks like, a preparation plan, how to choose a course, and how AI-103 compares with AB-731 and other credentials.

Who it is for

AI-103 targets developers and engineers who build AI applications and agents using Python and Microsoft Foundry. Microsoft recommends practical experience with Python, APIs and SDKs, and familiarity with AI, generative AI and Azure services.

It suits:

• Software developers adding generative AI features to applications.
• AI engineers building agents, retrieval-augmented generation (RAG) solutions and AI pipelines on Azure.
• Solution engineers and consultants who design and prototype AI solutions for clients.
• Data engineers moving into AI application work.

If you do not write code and your job is to decide where AI should be used, AB-731 is the right credential instead; see the comparison below.

Exam facts

Microsoft does not publish one global price on the exam page, so check the figure for your country at booking.

Skills measured

Microsoft's study guide (skills measured as of 16 April 2026) groups the exam into five areas:

The two largest areas together carry roughly 55–65% of the exam, so building and agent skills should take most of your practice time. Microsoft revises study guides, so check the date on the current version before you start.

1. Plan and manage an Azure AI solution (25–30%)

• Selecting the right Foundry services and models: large language models, small language models, multimodal models and tools.
• Planning resources, quotas and deployment options.
• Configuring solution components, security and access.
• Monitoring performance, cost and quality in production.

This area tests architectural judgement. Expect scenarios that ask which model or service fits a requirement, or how to deploy and monitor a solution responsibly.

2. Implement generative AI and agentic solutions (30–35%)

• Building generative AI applications.
• Working with prompts, system messages and model customisation.
• Building agents and orchestrating agent workflows.
• Connecting agents to tools, data and services.
• Applying safety, responsible AI and evaluation practices.

This is the heart of the exam. You should be able to build an agent that calls tools, reason about when to use an agent versus a simple prompt, and evaluate output quality.

3. Implement computer vision solutions (10–15%)

• Image and video analysis.
• Using Azure AI vision capabilities and relevant Foundry tools for vision workloads.

4. Implement text analysis solutions (10–15%)

• Entity and topic extraction, summarisation and structured JSON output.
• Sentiment, tone, safety and sensitive-content detection.
• Translation and domain-specific language workflows.
• Speech-to-text, text-to-speech and speech-enabled agent interactions.

5. Implement information extraction solutions (10–15%)

• Retrieval and grounding pipelines.
• Ingesting and indexing documents, images, audio and video.
• Semantic, hybrid and vector search.
• RAG ingestion, OCR and enrichment.
• Document extraction and structured output.

Key concepts to understand well

Retrieval-augmented generation (RAG)

RAG grounds a model's answers in your own data. A typical pipeline ingests documents, splits them into chunks, creates embeddings, stores them in a search index, retrieves relevant chunks for each question and passes them to the model with instructions to answer from that context. You should understand each step, how search choices (keyword, vector or hybrid) affect quality, and how to evaluate whether answers are grounded.

Agents and tools

An agent uses a model to decide which actions to take, such as calling an API, searching an index or running code, to complete a task. Expect to know how agents are given tools, how to constrain them, how to handle errors and how to orchestrate several steps or agents.

Evaluation and responsible AI

Microsoft weaves safety and evaluation through the skills. Understand how to measure quality (for example groundedness and relevance), how to detect and filter harmful content, how to protect sensitive data and how to monitor solutions after release.

Choosing models

Larger models are not always better. Smaller models can be cheaper and faster for well-defined tasks; multimodal models handle images and audio. The exam rewards choosing the right model for the requirement, including cost and latency.

Candidate profile

A typical successful candidate has built at least one end-to-end solution, not just followed tutorials: a retrieval-augmented app, an agent that calls tools, or a pipeline that extracts fields from documents. They know how to choose between model options, how to ground answers in data, and how to evaluate quality and safety. They are also comfortable reading Python SDK code, handling authentication to Azure services and reading error messages, because practical questions often hinge on those details.

A preparation plan

Adjust the pace to your experience. Experienced Azure developers may compress it; those new to Python or Azure should allow longer.

Useful resources

• Microsoft Learn learning paths mapped to AI-103.
• The AI-103 study guide, used as a checklist.
• Microsoft's official course, AI-103T00-A, which Microsoft lists as a four-day intermediate course.
• Practice assessments on Microsoft Learn, where available.
• Your own projects, which are the best preparation of all.

Managing Azure costs while you practise

• Use a dedicated subscription for practice.
• Set budget alerts.
• Delete deployments and search indexes after each session.
• Choose smaller models for experiments where possible.

What a preparation course needs

• Real code labs. The exam tests judgement about services and patterns, and that judgement comes from building.
• Foundry access. Ask how participants get an Azure environment, and who pays for consumption. Azure usage is usually separate from the training fee.
• Coverage of evaluation and responsible AI, woven through the labs rather than added at the end.
• Alignment with the current study guide. Check the date.
• A realistic duration. Microsoft's own instructor-led course, AI-103T00-A, runs for four days. Treat that as a reference point, not a promise about any provider's course.
• An instructor with real Azure AI development experience.

A worked scenario

Here is the kind of reasoning AI-103 asks for. It is an illustration written for this article, not an official exam question.

A legal team wants an assistant that answers questions about its contract library. Answers must cite the source clause, must not use information outside the library, and must keep client data within the organisation's Azure environment. Usage will be moderate, and the team wants to minimise development effort.

A strong answer would combine:

• Retrieval over the contract library with a search index, using hybrid search so both exact terms and meaning are matched.
• Grounded prompting that instructs the model to answer only from retrieved passages and to cite them.
• A managed model deployment in the organisation's Azure environment, chosen for adequate quality at moderate cost.
• Evaluation of groundedness and relevance before release, and content safety filtering.
• Access controls so users only retrieve documents they are allowed to see.

A weaker answer would fine-tune a model on the contracts. That adds effort and cost, does not guarantee citations and makes updates harder when contracts change. The scenario's constraints, citations and minimal effort, point to retrieval.

Common preparation mistakes

• Studying only the generative AI area and neglecting vision, text analysis and extraction, which together carry up to 45%.
• Following tutorials without building something of your own.
• Ignoring evaluation and safety, which appear across the skills.
• Using outdated material. Azure AI services and names change; check dates against the study guide.
• Leaving Azure resources running and paying for them.

Exam-day tips

• Read scenarios for constraints: cost, latency, data location, security and responsible AI requirements often decide the answer.
• Prefer managed services when the scenario does not require custom work.
• Watch for "minimise cost" or "minimise development effort" in the question; they change the right answer.
• Manage your time. With 120 minutes, leave time to review flagged questions and any interactive tasks.

AI-103 versus AB-731

Many organisations need both: leaders who can set direction and developers who can build it. Each individual normally needs only one. Our guide to AB-731 covers the leader track.

How AI-103 fits with other credentials

• Azure fundamentals exams can help if you are new to Azure.
• Data and cloud credentials build the infrastructure skills AI applications rely on; see our cloud certification paths.
• AI governance credentials cover policy and risk rather than development, and suit people who govern AI rather than build it.

Is AI-103 worth it?

AI-103 is worth it if you build, or are about to build, AI solutions on Azure, and the roles you want ask for Azure AI skills. Because it covers agents, retrieval and evaluation, it maps closely to the work organisations are doing now: internal assistants over company knowledge, document processing, customer support agents and AI features inside existing applications. It does not guarantee a role or pay rise; a portfolio of working solutions matters at least as much.

Get free certification pathway advice →

Key takeaways

• AI-103 leads to Microsoft Certified: Azure AI Apps and Agents Developer Associate.
• It is a 120-minute exam for Python developers building AI apps and agents with Microsoft Foundry.
• Generative AI and agentic solutions carry the most weight; vision, text and extraction together carry up to 45%.
• Hands-on building, including retrieval, agents and evaluation, is the best preparation.
• The exam price depends on your country; check it at booking.

Next step

CertScope's AI-103 training is being prepared with Python and Azure labs. If you want to be notified, or you need a team cohort, tell us your team's experience level.

View AI-103 Training Schedules →